An integrated machine learning framework for detecting and stratifying adverse drug reactions from bilingual clinical records

Medication-related harm remains a major source of preventable patient injury, and adverse drug reactions (ADRs) require effective pharmacovigilance systems. Although electronic medical records (EMRs) contain rich clinical information, ADR surveillance still relies heavily on manual chart review, which is labor-intensive and challenging in multilingual healthcare settings. This study introduces AIPharm, a two-component machine-learning framework for pharmacovigilance using bilingual Thai–English clinical narratives and structured EMR data. The first component detects ADRs from bilingual adverse event narratives, while the second stratifies ADR severity using structured clinical variables. These components address complementary but methodologically independent stages of an AI-assisted pharmacovigilance workflow. Using retrospective data from two psychiatric hospitals in Thailand between 2019 and 2025, the study analyzed 1278 bilingual narrative ADR reports and 4693 structured EMR records with 657 candidate clinical features. For narrative ADR detection, preprocessing ablation showed that numeric normalization was the most effective strategy, improving the weighted F1-score from 84.44% under the removal-based pipeline to 86.92%. Random Forest with numeric normalization achieved the strongest ADR detection performance while maintaining a low testing runtime of 14 ms. For structured ADR severity stratification, Random Forest achieved the highest test F1-score of 83.20% with 200 selected features. In contrast, Linear SVM achieved comparable performance with the fastest testing runtime, supporting its usefulness for rapid inference. Class-level analysis showed that Severe ADR remained the most difficult category to predict, with most misclassified severe cases assigned to the adjacent Moderate ADR class. AIPharm demonstrates the feasibility of combining bilingual clinical text processing with structured EMR-based modeling to support scalable, AI-assisted medication safety monitoring in multilingual healthcare environments. The framework may help prioritize suspected ADR reports, support severity assessment, and improve the consistency of pharmacovigilance workflows while preserving the role of expert clinical judgment.